Why LLM capability now looks like a funnel
Public understanding of LLM capability is becoming more uneven. In a long post on X, Andrej Karpathy argued that disagreement about what large language models can actually do is no longer just two groups talking past each other, but closer to a steep funnel of usage and access.
His rough estimate is that about 6 billion people have barely encountered LLMs, 1 to 2 billion use free ChatGPT-like products casually, about 20 million use frontier-grade models professionally in areas such as math and code, and roughly 5,000 people have access to the most advanced internal systems. These are napkin estimates, so they should be read as a framing device rather than precise statistics.
The gap comes from tasks, domains, and access
Karpathy’s central point is that experience shapes judgment. Free-tier users may see a sometimes awkward chatbot and treat it as a better search box or writing aid. Professional users working in code, math, and other digital, verifiable, economically valuable domains may instead see agents complete large projects that previously took weeks or months.
He attributes the gap to three forces: the impact grows with the ambition and time horizon of the task; model capability is jagged rather than uniform; and access differs across free, paid, and internal tiers. The same technology wave can therefore feel modest to consumers, dramatic to professional developers, and faster still to internal research teams.
The key open question is verification and governance
The post drew extensive discussion on Hacker News, which shows that disagreement itself has become part of AI adoption. Karpathy mentions internal capabilities such as long-running multi-agent collaboration, cyber offense and defense, scientific discovery, and mathematical progress, but the outside world has only limited previews and cannot independently verify every detail.
The important limitation is not simply whether AI is useful. It is how to evaluate capabilities seen at different access levels, how human review can keep pace with machine-generated work, and which breakthroughs can be publicly reproduced. If the understanding gap keeps widening, companies, policymakers, and the public may end up debating the same technology from very different factual baselines.